Can the network reveal
its own functional anatomy?
Do not assign regions first. Measure the effect of cutting every connection under three known training stimuli, discover groups of connections with similar causal fingerprints, commit predictions before test, then damage individual connections under unseen terminal demands.
What groups emerged?
Colored edges belong to the same statistically clustered response class, which may be disconnected in space. Thickness indicates recruited material. Click a connection or a group card to inspect it. Group labels are not predefined semantic concepts.
12 learned categories, no rectangular labels
Each edge's fingerprint records three actual half-strength intervention effects. Balanced clustering groups edges by those three numbers alone. All comparison methods use the same training observations.
Explain before seeing the effect
Every model predicts the loss increase from halving one edge's recruited conductance. The 49–51 test edges are selected from known trained material levels and fixed edge IDs—not test outcomes. The comparison includes material, training flow, an exact first-order gradient, fixed rectangles, and 20 additional shuffled null partitions.
Chosen model
Did the explanation predict anything new?
Exact finite cuts are evaluated on 12 disjoint, deterministic unseen source/sink pairs. Metrics are descriptive correlations and errors over a material-biased selection of edges, not independent biological samples.
| Training-only prediction | Pearson ↑ | RMSE ↓ | Top five cuts |
|---|
Held-out causal consistency
Within-group standard deviation of unseen individual-cut effects (smaller is more homogeneous). Different groupings share identical edges and damage experiments.
20 shuffled explanations
What counts as discovery?
A new group is useful only if its training-induced interpretation predicts unobserved causal changes better than equally informed simpler explanations. Its visible shape or cluster coherence is not sufficient.
Known toy dynamics
All training and test stimuli live in one predetermined Kirchhoff lattice, with a fixed number of possible edges. The clustering sees three supervised training probes; its category names are retrospective summaries. This does not replicate Goodfire's learned neural representations.